[Beowulf] RE: Capitalization Rates - How often should you replace a cluster? (resent - 1st sending wasn't posted ).

[Robert G. Brown]

On Fri, 16 Jan 2009, Greg Lindahl wrote:

> On Thu, Jan 15, 2009 at 04:26:00PM -0500, Lechner, David A. wrote:
>
>> Sometimes the newer technology uses less power and is cheaper to
>> operate....(anyone ever create a KW/MFLOP vs. Time curve?  has that
>> really gone down? )
>
> I'd think that you would have looked that up before you posted! I
> don't have the curve, but it's falling rapidly enough that it's
> obvious that an older cluster can be replaced with a mch smaller, new
> cluster of equal performance for the cost of power. How old?  I don't
> know, but I'd guess it's less than 6 years. And anyone who rents
> machine room space probably replaces in ~ 3.

I've done the estimate in detail several times (estimate because it
depends a bit on power costs in your area etc) and I get between years 4
and 5.  I'm sure the estimates are described in detail in the archives
-- probably more than once.

If one factors in additional costs -- maintenance is the usual one, but
machine room costs outside of power is another valid one for sure --
more rapid replacement is indicated.

I've had discussions on and off line with people who argue that to
minimize "TCO" replacement every twelve months is in order.  In general
I don't agree with that -- maybe in a horribly inefficient corporate
environment with a high payoff to staying bleeding edge current and
equally high expenses for management and maintenance, but noplace sane
-- and would "recommend" a general replacement cycle of the period you
keep your systems under HW maintenance.  Usually three or four years, in
other words.

> Since many clusters are funded by one-off funding sources, and often
> the cluster buyer doesn't have to directly pay for power or rent,
> you'll probably find that actual replacement dates don't make any
> sense...  not only for academic sites, but commercial ones, too.

Agreed.  If you get money this year for a cluster, you might run it for
four years (under maintenance) and then just use it until it dies
(sometimes canniblizing along the way to the last node standing) if you
can't get new funding for a replacement, especially if your University
pays for the power and human management out of overhead etc.

    rgb

>
> -- greg
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>

Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977